Overview

In early 2026 Amazon upgraded its search engine, replacing simple keyword matching with a semantic system driven by the COSMO algorithm and Rufus AI. The new engine treats each product as a node in a knowledge graph, linking listings through attributes, customer Q&A, review sentiment and visual cues. Sellers who continue to rely on keyword stuffing will see reduced visibility and must shift to Knowledge Graph Optimization (KGO).

Key Points

• Semantic over lexical search — COSMO evaluates the relationships between items, categories and shopper intent instead of looking for exact keyword strings.
• Products as knowledge nodes — Every listing now connects to other products via structured attributes, Q&A content, review tone and image data, forming an inter‑linked graph rather than an isolated page.
• Multimodal content processing — The AI reads text from images with OCR, interprets video frames with computer vision, and treats A+ modules as technical documentation for indexing.
• Conversational query matching — Rufus translates natural‑language questions into attribute matches, allowing queries like “Can this blender crush ice?” to surface the correct ASIN without the phrase appearing in the title.
• Data consistency rewarded — Listings that present uniform, verifiable information across all fields are ranked higher; contradictory or missing data triggers suppression in search results.

How Rufus AI and COSMO Operate

1. Attribute ingestion — When a seller uploads a flat‑file or edits a product, COSMO extracts every backend field (material, size, target audience, etc.) and maps it to Amazon’s taxonomy. For example, a “stainless‑steel” entry placed in the “Material” field is linked to the “Stainless Steel” node in the knowledge graph, instantly connecting the product to related queries about durability.
2. Multimodal enrichment — Rufus scans the main image, alt text, and any video, pulling out visual descriptors such as “safety‑lock blade” or “LED display”. An infographic that lists “120 W motor” in a clear font is read via OCR and added as a factual attribute, allowing the system to answer “What is the motor power?” even if the phrase never appears in the copy.